Spaces:
Runtime error
Runtime error
Stable-X
commited on
Commit
Β·
b353dc0
1
Parent(s):
35c32ba
Update scheduler
Browse files
stablenormal/scheduler/heuristics_ddimsampler.py
ADDED
|
@@ -0,0 +1,236 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import math
|
| 2 |
+
from dataclasses import dataclass
|
| 3 |
+
from typing import List, Optional, Tuple, Union
|
| 4 |
+
|
| 5 |
+
import numpy as np
|
| 6 |
+
import torch
|
| 7 |
+
from diffusers.schedulers.scheduling_ddim import DDIMSchedulerOutput, DDIMScheduler
|
| 8 |
+
from diffusers.schedulers.scheduling_utils import SchedulerMixin
|
| 9 |
+
from diffusers.configuration_utils import register_to_config, ConfigMixin
|
| 10 |
+
import pdb
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
class HEURI_DDIMScheduler(DDIMScheduler, SchedulerMixin, ConfigMixin):
|
| 14 |
+
|
| 15 |
+
def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
|
| 16 |
+
"""
|
| 17 |
+
Sets the discrete timesteps used for the diffusion chain (to be run before inference).
|
| 18 |
+
|
| 19 |
+
Args:
|
| 20 |
+
num_inference_steps (`int`):
|
| 21 |
+
The number of diffusion steps used when generating samples with a pre-trained model.
|
| 22 |
+
"""
|
| 23 |
+
|
| 24 |
+
if num_inference_steps > self.config.num_train_timesteps:
|
| 25 |
+
raise ValueError(
|
| 26 |
+
f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
|
| 27 |
+
f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
|
| 28 |
+
f" maximal {self.config.num_train_timesteps} timesteps."
|
| 29 |
+
)
|
| 30 |
+
|
| 31 |
+
self.num_inference_steps = num_inference_steps
|
| 32 |
+
|
| 33 |
+
# "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
|
| 34 |
+
if self.config.timestep_spacing == "linspace":
|
| 35 |
+
timesteps = (
|
| 36 |
+
np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
|
| 37 |
+
.round()[::-1]
|
| 38 |
+
.copy()
|
| 39 |
+
.astype(np.int64)
|
| 40 |
+
)
|
| 41 |
+
elif self.config.timestep_spacing == "leading":
|
| 42 |
+
step_ratio = self.config.num_train_timesteps // self.num_inference_steps
|
| 43 |
+
# creates integer timesteps by multiplying by ratio
|
| 44 |
+
# casting to int to avoid issues when num_inference_step is power of 3
|
| 45 |
+
timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
|
| 46 |
+
timesteps += self.config.steps_offset
|
| 47 |
+
elif self.config.timestep_spacing == "trailing":
|
| 48 |
+
step_ratio = self.config.num_train_timesteps / self.num_inference_steps
|
| 49 |
+
# creates integer timesteps by multiplying by ratio
|
| 50 |
+
# casting to int to avoid issues when num_inference_step is power of 3
|
| 51 |
+
timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64)
|
| 52 |
+
timesteps -= 1
|
| 53 |
+
else:
|
| 54 |
+
raise ValueError(
|
| 55 |
+
f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'."
|
| 56 |
+
)
|
| 57 |
+
|
| 58 |
+
timesteps = torch.from_numpy(timesteps).to(device)
|
| 59 |
+
naive_sampling_step = num_inference_steps //2
|
| 60 |
+
|
| 61 |
+
self.naive_sampling_step = naive_sampling_step
|
| 62 |
+
|
| 63 |
+
timesteps[:naive_sampling_step] = timesteps[naive_sampling_step] # refine on step 5 for 5 steps, then backward from step 6
|
| 64 |
+
|
| 65 |
+
timesteps = [timestep + 1 for timestep in timesteps]
|
| 66 |
+
|
| 67 |
+
self.timesteps = timesteps
|
| 68 |
+
self.gap = self.config.num_train_timesteps // self.num_inference_steps
|
| 69 |
+
self.prev_timesteps = [timestep for timestep in self.timesteps[1:]]
|
| 70 |
+
self.prev_timesteps.append(torch.zeros_like(self.prev_timesteps[-1]))
|
| 71 |
+
|
| 72 |
+
def step(
|
| 73 |
+
self,
|
| 74 |
+
model_output: torch.Tensor,
|
| 75 |
+
timestep: int,
|
| 76 |
+
prev_timestep: int,
|
| 77 |
+
sample: torch.Tensor,
|
| 78 |
+
eta: float = 0.0,
|
| 79 |
+
use_clipped_model_output: bool = False,
|
| 80 |
+
generator=None,
|
| 81 |
+
cur_step=None,
|
| 82 |
+
gauss_latent=None,
|
| 83 |
+
variance_noise: Optional[torch.Tensor] = None,
|
| 84 |
+
return_dict: bool = True,
|
| 85 |
+
) -> Union[DDIMSchedulerOutput, Tuple]:
|
| 86 |
+
"""
|
| 87 |
+
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
|
| 88 |
+
process from the learned model outputs (most often the predicted noise).
|
| 89 |
+
|
| 90 |
+
Args:
|
| 91 |
+
model_output (`torch.Tensor`):
|
| 92 |
+
The direct output from learned diffusion model.
|
| 93 |
+
timestep (`float`):
|
| 94 |
+
The current discrete timestep in the diffusion chain.
|
| 95 |
+
pre_timestep (`float`):
|
| 96 |
+
next_timestep
|
| 97 |
+
sample (`torch.Tensor`):
|
| 98 |
+
A current instance of a sample created by the diffusion process.
|
| 99 |
+
eta (`float`):
|
| 100 |
+
The weight of noise for added noise in diffusion step.
|
| 101 |
+
use_clipped_model_output (`bool`, defaults to `False`):
|
| 102 |
+
If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary
|
| 103 |
+
because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no
|
| 104 |
+
clipping has happened, "corrected" `model_output` would coincide with the one provided as input and
|
| 105 |
+
`use_clipped_model_output` has no effect.
|
| 106 |
+
generator (`torch.Generator`, *optional*):
|
| 107 |
+
A random number generator.
|
| 108 |
+
variance_noise (`torch.Tensor`):
|
| 109 |
+
Alternative to generating noise with `generator` by directly providing the noise for the variance
|
| 110 |
+
itself. Useful for methods such as [`CycleDiffusion`].
|
| 111 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
| 112 |
+
Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`.
|
| 113 |
+
|
| 114 |
+
Returns:
|
| 115 |
+
[`~schedulers.scheduling_utils.DDIMSchedulerOutput`] or `tuple`:
|
| 116 |
+
If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a
|
| 117 |
+
tuple is returned where the first element is the sample tensor.
|
| 118 |
+
|
| 119 |
+
"""
|
| 120 |
+
if self.num_inference_steps is None:
|
| 121 |
+
raise ValueError(
|
| 122 |
+
"Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
|
| 123 |
+
)
|
| 124 |
+
|
| 125 |
+
# See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
|
| 126 |
+
# Ideally, read DDIM paper in-detail understanding
|
| 127 |
+
|
| 128 |
+
# Notation (<variable name> -> <name in paper>
|
| 129 |
+
# - pred_noise_t -> e_theta(x_t, t)
|
| 130 |
+
# - pred_original_sample -> f_theta(x_t, t) or x_0
|
| 131 |
+
# - std_dev_t -> sigma_t
|
| 132 |
+
# - eta -> Ξ·
|
| 133 |
+
# - pred_sample_direction -> "direction pointing to x_t"
|
| 134 |
+
# - pred_prev_sample -> "x_t-1"
|
| 135 |
+
|
| 136 |
+
# 1. get previous step value (=t-1)
|
| 137 |
+
# trick from heuri_sampling
|
| 138 |
+
if cur_step == self.naive_sampling_step and timestep == prev_timestep:
|
| 139 |
+
timestep += self.gap
|
| 140 |
+
|
| 141 |
+
prev_timestep = prev_timestep # NOTE naive sampling
|
| 142 |
+
|
| 143 |
+
# 2. compute alphas, betas
|
| 144 |
+
alpha_prod_t = self.alphas_cumprod[timestep]
|
| 145 |
+
alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
|
| 146 |
+
|
| 147 |
+
beta_prod_t = 1 - alpha_prod_t
|
| 148 |
+
|
| 149 |
+
# 3. compute predicted original sample from predicted noise also called
|
| 150 |
+
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
| 151 |
+
if self.config.prediction_type == "epsilon":
|
| 152 |
+
pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
|
| 153 |
+
pred_epsilon = model_output
|
| 154 |
+
elif self.config.prediction_type == "sample":
|
| 155 |
+
pred_original_sample = model_output
|
| 156 |
+
pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
|
| 157 |
+
elif self.config.prediction_type == "v_prediction":
|
| 158 |
+
pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
|
| 159 |
+
pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
|
| 160 |
+
else:
|
| 161 |
+
raise ValueError(
|
| 162 |
+
f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
|
| 163 |
+
" `v_prediction`"
|
| 164 |
+
)
|
| 165 |
+
|
| 166 |
+
# 4. Clip or threshold "predicted x_0"
|
| 167 |
+
if self.config.thresholding:
|
| 168 |
+
pred_original_sample = self._threshold_sample(pred_original_sample)
|
| 169 |
+
|
| 170 |
+
# 5. compute variance: "sigma_t(Ξ·)" -> see formula (16)
|
| 171 |
+
# Ο_t = sqrt((1 β Ξ±_tβ1)/(1 β Ξ±_t)) * sqrt(1 β Ξ±_t/Ξ±_tβ1)
|
| 172 |
+
variance = self._get_variance(timestep, prev_timestep)
|
| 173 |
+
std_dev_t = eta * variance ** (0.5)
|
| 174 |
+
|
| 175 |
+
if use_clipped_model_output:
|
| 176 |
+
# the pred_epsilon is always re-derived from the clipped x_0 in Glide
|
| 177 |
+
pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
|
| 178 |
+
|
| 179 |
+
# 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
| 180 |
+
pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
|
| 181 |
+
|
| 182 |
+
# 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
| 183 |
+
if gauss_latent == None:
|
| 184 |
+
gauss_latent = torch.randn_like(pred_original_sample)
|
| 185 |
+
prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
|
| 186 |
+
|
| 187 |
+
if eta > 0:
|
| 188 |
+
if variance_noise is not None and generator is not None:
|
| 189 |
+
raise ValueError(
|
| 190 |
+
"Cannot pass both generator and variance_noise. Please make sure that either `generator` or"
|
| 191 |
+
" `variance_noise` stays `None`."
|
| 192 |
+
)
|
| 193 |
+
|
| 194 |
+
if variance_noise is None:
|
| 195 |
+
variance_noise = randn_tensor(
|
| 196 |
+
model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
|
| 197 |
+
)
|
| 198 |
+
variance = std_dev_t * variance_noise
|
| 199 |
+
|
| 200 |
+
prev_sample = prev_sample + variance
|
| 201 |
+
|
| 202 |
+
if cur_step < self.naive_sampling_step:
|
| 203 |
+
prev_sample = self.add_noise(pred_original_sample, gauss_latent, timestep)
|
| 204 |
+
|
| 205 |
+
if not return_dict:
|
| 206 |
+
return (prev_sample,)
|
| 207 |
+
|
| 208 |
+
return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
|
| 209 |
+
|
| 210 |
+
|
| 211 |
+
|
| 212 |
+
def add_noise(
|
| 213 |
+
self,
|
| 214 |
+
original_samples: torch.Tensor,
|
| 215 |
+
noise: torch.Tensor,
|
| 216 |
+
timesteps: torch.IntTensor,
|
| 217 |
+
) -> torch.Tensor:
|
| 218 |
+
# Make sure alphas_cumprod and timestep have same device and dtype as original_samples
|
| 219 |
+
# Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
|
| 220 |
+
# for the subsequent add_noise calls
|
| 221 |
+
self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
|
| 222 |
+
alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
|
| 223 |
+
timesteps = timesteps.to(original_samples.device)
|
| 224 |
+
|
| 225 |
+
sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
|
| 226 |
+
sqrt_alpha_prod = sqrt_alpha_prod.flatten()
|
| 227 |
+
while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
|
| 228 |
+
sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
|
| 229 |
+
|
| 230 |
+
sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
|
| 231 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
|
| 232 |
+
while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
|
| 233 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
|
| 234 |
+
|
| 235 |
+
noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
|
| 236 |
+
return noisy_samples
|